---

The QNX Neutrino Microkernel

• Introduction
• The implementation of QNX Neutrino
• System services
• Threads and processes
• Thread scheduling
• Synchronization services
• QNX Neutrino IPC
• Clock and timer services
• Interrupt handling

Introduction

The QNX Neutrino microkernel, procnto, implements the core POSIX features used in embedded realtime systems, along with the fundamental QNX Neutrino message-passing services. The POSIX features that aren't implemented in the microkernel (file and device I/O, for example) are provided by optional processes and shared libraries.

Successive QNX microkernels have seen a reduction in the code required to implement a given kernel call. The object definitions at the lowest layer in the kernel code have become more specific, allowing greater code reuse (such as folding various forms of POSIX signals, realtime signals, and QNX pulses into common data structures and code to manipulate those structures).

At its lowest level, the microkernel contains a few fundamental objects and the highly tuned routines that manipulate them. The OS is built from this foundation.

---

---

The QNX Neutrino microkernel.

Some developers have assumed that our microkernel is implemented entirely in assembly code for size or performance reasons. In fact, our implementation is coded primarily in C; size and performance goals are achieved through successively refined algorithms and data structures, rather than via assembly-level peep-hole optimizations.

The implementation of QNX Neutrino

Historically, the "application pressure" on QNX operating systems has been from both ends of the computing spectrum -- from memory-limited embedded systems all the way up to high-end SMP (symmetrical multiprocessing) machines with gigabytes of physical memory. Accordingly, the design goals for QNX Neutrino accommodate both seemingly exclusive sets of functionality. Pursuing these goals is intended to extend the reach of systems well beyond what other OS implementations could address.

POSIX realtime and thread extensions

Since QNX Neutrino implements the majority of the realtime and thread services directly in the microkernel, these services are available even without the presence of additional OS modules.

In addition, some of the profiles defined by POSIX suggest that these services be present without necessarily requiring a process model. In order to accommodate this, the OS provides direct support for threads, but relies on its process manager portion to extend this functionality to processes containing multiple threads.

Note that many realtime executives and kernels provide only a nonmemory-protected threaded model, with no process model and/or protected memory model at all. Without a process model, full POSIX compliance cannot be achieved.

System services

The QNX Neutrino microkernel has kernel calls to support the following:

• threads
• message passing
• signals
• clocks
• timers
• interrupt handlers
• semaphores
• mutual exclusion locks (mutexes)
• condition variables (condvars)
• barriers.

The entire OS is built upon these calls. The OS is fully preemptible, even while passing messages between processes; it resumes the message pass where it left off before preemption.

The minimal complexity of the microkernel helps place an upper bound on the longest nonpreemptible code path through the kernel, while the small code size makes addressing complex multiprocessor issues a tractable problem. Services were chosen for inclusion in the microkernel on the basis of having a short execution path. Operations requiring significant work (e.g. process loading) were assigned to external processes/threads, where the effort to enter the context of that thread would be insignificant compared to the work done within the thread to service the request.

Rigorous application of this rule to dividing the functionality between the kernel and external processes destroys the myth that a microkernel OS must incur higher runtime overhead than a monolithic kernel OS. Given the work done between context switches (implicit in a message pass), and the very quick context-switch times that result from the simplified kernel, the time spent performing context switches becomes "lost in the noise" of the work done to service the requests communicated by the message passing between the processes that make up the OS.

The following diagram shows the preemption details for the non-SMP kernel (x86 implementation).

---

---

QNX Neutrino preemption details.

Interrupts are disabled, or preemption is held off, for only very brief intervals (typically in the order of hundreds of nanoseconds).

Threads and processes

When building an application (realtime, embedded, graphical, or otherwise), the developer may want several algorithms within the application to execute concurrently. This concurrency is achieved by using the POSIX thread model, which defines a process as containing one or more threads of execution.

A thread can be thought of as the minimum "unit of execution," the unit of scheduling and execution in the microkernel. A process, on the other hand, can be thought of as a "container" for threads, defining the "address space" within which threads will execute. A process will always contain at least one thread.

Depending on the nature of the application, threads might execute independently with no need to communicate between the algorithms (unlikely), or they may need to be tightly coupled, with high-bandwidth communications and tight synchronization. To assist in this communication and synchronization, QNX Neutrino provides a rich variety of IPC and synchronization services.

The following pthread_* (POSIX Threads) library calls don't involve any microkernel thread calls:

• pthread_attr_destroy()
• pthread_attr_getdetachstate()
• pthread_attr_getinheritsched()
• pthread_attr_getschedparam()
• pthread_attr_getschedpolicy()
• pthread_attr_getscope()
• pthread_attr_getstackaddr()
• pthread_attr_getstacksize()
• pthread_attr_init()
• pthread_attr_setdetachstate()
• pthread_attr_setinheritsched()
• pthread_attr_setschedparam()
• pthread_attr_setschedpolicy()
• pthread_attr_setscope()
• pthread_attr_setstackaddr()
• pthread_attr_setstacksize()
• pthread_cleanup_pop()
• pthread_cleanup_push()
• pthread_equal()
• pthread_getspecific()
• pthread_setspecific()
• pthread_testcancel()
• pthread_key_create()
• pthread_key_delete()
• pthread_once()
• pthread_self()
• pthread_setcancelstate()
• pthread_setcanceltype()

The following table lists the POSIX thread calls that have a corresponding microkernel thread call, allowing you to choose either interface:

POSIX call	Microkernel call	Description
pthread_create()	ThreadCreate()	Create a new thread of execution.
pthread_exit()	ThreadDestroy()	Destroy a thread.
pthread_detach()	ThreadDetach()	Detach a thread so it doesn't need to be joined.
pthread_join()	ThreadJoin()	Join a thread waiting for its exit status.
pthread_cancel()	ThreadCancel()	Cancel a thread at the next cancellation point.
N/A	ThreadCtl()	Change a thread's Neutrino-specific thread characteristics.
pthread_mutex_init()	SyncTypeCreate()	Create a mutex.
pthread_mutex_destroy()	SyncDestroy()	Destroy a mutex.
pthread_mutex_lock()	SyncMutexLock()	Lock a mutex.
pthread_mutex_trylock()	SyncMutexLock()	Conditionally lock a mutex.
pthread_mutex_unlock()	SyncMutexUnlock()	Unlock a mutex.
pthread_cond_init()	SyncTypeCreate()	Create a condition variable.
pthread_cond_destroy()	SyncDestroy()	Destroy a condition variable.
pthread_cond_wait()	SyncCondvarWait()	Wait on a condition variable.
pthread_cond_signal()	SyncCondvarSignal()	Signal a condition variable.
pthread_cond_broadcast()	SyncCondvarSignal()	Broadcast a condition variable.
pthread_getschedparam()	SchedGet()	Get scheduling parameters and policy of thread.
pthread_setschedparam()	SchedSet()	Set scheduling parameters and policy of thread.
pthread_sigmask()	SignalProcmask()	Examine or set a thread's signal mask.
pthread_kill()	SignalKill()	Send a signal to a specific thread.

The OS can be configured to provide a mix of threads and processes (as defined by POSIX). Each process is MMU-protected from each other, and each process may contain one or more threads that share the process's address space.

The environment you choose affects not only the concurrency capabilities of the application, but also the IPC and synchronization services the application might make use of.

Thread attributes

Although threads within a process share everything within the process's address space, each thread still has some "private" data. In some cases, this private data is protected within the kernel (e.g. the tid or thread ID), while other private data resides unprotected in the process's address space (e.g. each thread has a stack for its own use). Some of the more noteworthy thread-private resources are:

tid

Each thread is identified by an integer thread ID, starting at 1. The tid is unique within the thread's process.

name

Starting with the QNX Neutrino Core OS 6.3.2, you can assign a name to a thread; see the entries for pthread_getname_np() and pthread_setname_np() in the Neutrino Library Reference. Utilities such as dumper and pidin support thread names. Thread names are a Neutrino extension.

register set

Each thread has its own instruction pointer (IP), stack pointer (SP), and other processor-specific register context.

stack

Each thread executes on its own stack, stored within the address space of its process.

signal mask

Each thread has its own signal mask.

thread local storage

A thread has a system-defined data area called "thread local storage" (TLS). The TLS is used to store "per-thread" information (such as tid, pid, stack base, errno, and thread-specific key/data bindings). The TLS doesn't need to be accessed directly by a user application. A thread can have user-defined data associated with a thread-specific data key.

cancellation handlers

Callback functions that are executed when the thread terminates.

Thread-specific data, implemented in the pthread library and stored in the TLS, provides a mechanism for associating a process global integer key with a unique per-thread data value. To use thread-specific data, you first create a new key and then bind a unique data value to the key (per thread). The data value may, for example, be an integer or a pointer to a dynamically allocated data structure. Subsequently, the key can return the bound data value per thread.

A typical application of thread-specific data is for a thread-safe function that needs to maintain a context for each calling thread.

---

---

Sparse matrix (tid,key) to value mapping.

You use the following functions to create and manipulate this data:

Function	Description
pthread_key_create()	Create a data key with destructor function
pthread_key_delete()	Destroy a data key
pthread_setspecific()	Bind a data value to a data key
pthread_getspecific()	Return the data value bound to a data key

Thread life cycle

The number of threads within a process can vary widely, with threads being created and destroyed dynamically. Thread creation (pthread_create()) involves allocating and initializing the necessary resources within the process's address space (e.g. thread stack) and starting the execution of the thread at some function in the address space.

Thread termination (pthread_exit(), pthread_cancel()) involves stopping the thread and reclaiming the thread's resources. As a thread executes, its state can generally be described as either "ready" or "blocked." More specifically, it can be one of the following:

---

---

Possible thread states.

CONDVAR

The thread is blocked on a condition variable (e.g. it called pthread_cond_wait()).

DEAD

The thread has terminated and is waiting for a join by another thread.

INTERRUPT

The thread is blocked waiting for an interrupt (i.e. it called InterruptWait()).

JOIN

The thread is blocked waiting to join another thread (e.g. it called pthread_join()).

MUTEX

The thread is blocked on a mutual exclusion lock (e.g. it called pthread_mutex_lock()).

NANOSLEEP

The thread is sleeping for a short time interval (e.g. it called nanosleep()).

NET_REPLY

The thread is waiting for a reply to be delivered across the network (i.e. it called MsgReply*()).

NET_SEND

The thread is waiting for a pulse or signal to be delivered across the network (i.e. it called MsgSendPulse(), MsgDeliverEvent(), or SignalKill()).

READY

The thread is waiting to be executed while the processor executes another thread of equal or higher priority.

RECEIVE

The thread is blocked on a message receive (e.g. it called MsgReceive()).

REPLY

The thread is blocked on a message reply (i.e. it called MsgSend(), and the server received the message).

RUNNING

The thread is being executed by a processor.

SEM

The thread is waiting for a semaphore to be posted (i.e. it called SyncSemWait()).

SEND

The thread is blocked on a message send (e.g. it called MsgSend(), but the server hasn't yet received the message).

SIGSUSPEND

The thread is blocked waiting for a signal (i.e. it called sigsuspend()).

SIGWAITINFO

The thread is blocked waiting for a signal (i.e. it called sigwaitinfo()).

STACK

The thread is waiting for the virtual address space to be allocated for the thread's stack (parent will have called ThreadCreate()).

STOPPED

The thread is blocked waiting for a SIGCONT signal.

WAITCTX

The thread is waiting for a noninteger (e.g. floating point) context to become available for use.

WAITPAGE

The thread is waiting for physical memory to be allocated for a virtual address.

WAITTHREAD

The thread is waiting for a child thread to finish creating itself (i.e. it called ThreadCreate()).

Thread scheduling

When scheduling decisions are made

The execution of a running thread is temporarily suspended whenever the microkernel is entered as the result of a kernel call, exception, or hardware interrupt. A scheduling decision is made whenever the execution state of any thread changes -- it doesn't matter which processes the threads might reside within. Threads are scheduled globally across all processes.

Normally, the execution of the suspended thread will resume, but the scheduler will perform a context switch from one thread to another whenever the running thread:

• is blocked
• is preempted
• yields.

When thread is blocked

The running thread blocks when it must wait for some event to occur (response to an IPC request, wait on a mutex, etc.). The blocked thread is removed from the ready queue and the highest-priority ready thread is then run. When the blocked thread is subsequently unblocked, it is placed on the end of the ready queue for that priority level.

When thread is preempted

The running thread is preempted when a higher-priority thread is placed on the ready queue (it becomes READY, as the result of its block condition being resolved). The preempted thread remains at the beginning of the ready queue for that priority and the higher-priority thread runs.

When thread yields

The running thread voluntarily yields the processor (sched_yield()) and is placed on the end of the ready queue for that priority. The highest-priority thread then runs (which may still be the thread that just yielded).

Scheduling priority

Every thread is assigned a priority. The scheduler selects the next thread to run by looking at the priority assigned to every thread that is READY (i.e. capable of using the CPU). The thread with the highest priority is selected to run.

The following diagram shows the ready queue for six threads (A-F) that are READY. All other threads (G-Z) are BLOCKED. Thread A is currently running. Thread A, B, and C are at the highest priority, so they'll share the processor based on the running thread's scheduling algorithm.

---

---

The ready queue.

The OS supports a total of 256 scheduling priority levels. A non-root thread can set its priority to a level from 1 to 63 (the highest priority), independent of the scheduling policy. Only root threads (i.e. those whose effective uid is 0) are allowed to set priorities above 63. The special idle thread (in the process manager) has priority 0 and is always ready to run. A thread inherits the priority of its parent thread by default.

You can change the allowed priority range for non-root processes with the procnto -P option:

procnto -P priority

Here's a summary of the ranges:

Priority level	Owner
0	Idle thread
1 through priority - 1	Non-root or root
priority through 255	root

Note that in order to prevent priority inversion, the kernel may temporarily boost a thread's priority. For more information, see "Priority inheritance and mutexes" and "Priority inheritance and messages," later in this chapter.

The threads on the ready queue are ordered by priority. The ready queue is actually implemented as 256 separate queues, one for each priority. Most of the time, threads are queued in FIFO order in the queue of their priority. (The exception is a server thread that's coming out of a RECEIVE-blocked state with a message from a client; the server thread is inserted at the head of the queue for that priority -- that is, the order is LIFO, not FIFO.) The first thread in the highest-priority queue is selected to run.

Scheduling algorithms

To meet the needs of various applications, QNX Neutrino provides these scheduling algorithms:

• FIFO scheduling
• round-robin scheduling
• sporadic scheduling.

Each thread in the system may run using any method. The methods are effective on a per-thread basis, not on a global basis for all threads and processes on a node.

Remember that the FIFO and round-robin scheduling algorithms apply only when two or more threads that share the same priority are READY (i.e. the threads are directly competing with each other). The sporadic method, however, employs a "budget" for a thread's execution. In all cases, if a higher-priority thread becomes READY, it immediately preempts all lower-priority threads.

In the following diagram, three threads of equal priority are READY. If Thread A blocks, Thread B will run.

---

---

Thread A blocks, Thread B runs.

Although a thread inherits its scheduling algorithm from its parent process, the thread can request to change the algorithm applied by the kernel.

FIFO scheduling

In FIFO scheduling, a thread selected to run continues executing until it:

• voluntarily relinquishes control (e.g. it blocks)
• is preempted by a higher-priority thread.

---

---

FIFO scheduling.

Round-robin scheduling

In round-robin scheduling, a thread selected to run continues executing until it:

• voluntarily relinquishes control
• is preempted by a higher-priority thread
• consumes its timeslice.

As the following diagram shows, Thread A ran until it consumed its timeslice; the next READY thread (Thread B) now runs:

---

---

Round-robin scheduling.

A timeslice is the unit of time assigned to every process. Once it consumes its timeslice, a thread is preempted and the next READY thread at the same priority level is given control. A timeslice is 4 * the clock period. (For more information, see the entry for ClockPeriod() in the Library Reference.)

Sporadic scheduling

The sporadic scheduling algorithm is generally used to provide a capped limit on the execution time of a thread within a given period of time. This behavior is essential when Rate Monotonic Analysis (RMA) is being performed on a system that services both periodic and aperiodic events. Essentially, this algorithm allows a thread to service aperiodic events without jeopardizing the hard deadlines of other threads or processes in the system.

As in FIFO scheduling, a thread using sporadic scheduling continues executing until it blocks or is preempted by a higher-priority thread. And as in adaptive scheduling, a thread using sporadic scheduling will drop in priority, but with sporadic scheduling you have much more precise control over the thread's behavior.

Under sporadic scheduling, a thread's priority can oscillate dynamically between a foreground or normal priority and a background or low priority. Using the following parameters, you can control the conditions of this sporadic shift:

Initial budget (C)

The amount of time a thread is allowed to execute at its normal priority (N) before being dropped to its low priority (L).

Low priority (L)

The priority level to which the thread will drop. The thread executes at this lower priority (L) while in the background, and runs at normal priority (N) while in the foreground.

Replenishment period (T)

The period of time during which a thread is allowed to consume its execution budget. To schedule replenishment operations, the POSIX implementation also uses this value as the offset from the time the thread becomes READY.

Max number of pending replenishments

This value limits the number of replenishment operations that can take place, thereby bounding the amount of system overhead consumed by the sporadic scheduling policy.

---

In a poorly configured system, a thread's execution budget may become eroded because of too much blocking -- i.e. it won't receive enough replenishments.

---

As the following diagram shows, the sporadic scheduling policy establishes a thread's initial execution budget (C), which is consumed by the thread as it runs and is replenished periodically (for the amount T). When a thread blocks, the amount of the execution budget that's been consumed (R) is arranged to be replenished at some later time (e.g. at 40 msec) after the thread first became ready to run.

---

---

A thread's budget is replenished periodically.

At its normal priority N, a thread will execute for the amount of time defined by its initial execution budget C. As soon as this time is exhausted, the priority of the thread will drop to its low priority L until the replenishment operation occurs.

Assume, for example, a system where the thread never blocks or is never preempted:

---

---

Thread drops in priority until its budget is replenished.

Here the thread will drop to its low-priority (background) level, where it may or may not get a chance to run depending on the priority of other threads in the system.

Once the replenishment occurs, the thread's priority is raised to its original level. This guarantees that within a properly configured system, the thread will be given the opportunity every period T to run for a maximum execution time C. This ensures that a thread running at priority N will consume only C/T percent of the system's resources.

When a thread blocks multiple times, then several replenishment operations may be started and occur at different times. This could mean that the thread's execution budget will total C within a period T; however, the execution budget may not be contiguous during that period.

---

---

Thread oscillates between high and low priority.

In the diagram above, the thread has a budget (C) of 10 msec to be consumed within each 40-msec replenishment period (T).

1. The initial run of the thread is blocked after 3 msec, so a replenishment operation of 3 msec is scheduled to begin at the 40-msec mark, i.e. when its first replenishment period has elapsed.
2. The thread gets an opportunity to run again at 6 msec, which marks the start of another replenishment period (T). The thread still has 7 msec remaining in its budget.
3. The thread runs without blocking for 7 msec, thereby exhausting its budget, and then drops to low priority L, where it may or may not be able to execute. A replenishment of 7 msec is scheduled to occur at 46 msec (40 + 6), i.e. when T has elapsed.
4. The thread has 3 msec of its budget replenished at 40 msec (see Step 1) and is therefore boosted back to its normal priority.
5. The thread consumes the 3 msec of its budget and then is dropped back to the low priority.
6. The thread has 7 msec of its budget replenished at 46 msec (see Step 3) and is boosted back to its normal priority.

And so on. The thread will continue to oscillate between its two priority levels, servicing aperiodic events in your system in a controlled, predictable manner.

Manipulating priority and scheduling algorithms

A thread's priority can vary during its execution, either from direct manipulation by the thread itself or from the kernel adjusting the thread's priority as it receives a message from a higher-priority thread.

In addition to priority, you can also select the scheduling algorithm that the kernel will use for the thread. Here are the POSIX calls for performing these manipulations, along with the microkernel calls used by these library routines:

POSIX call	Microkernel call	Description
sched_getparam()	SchedGet()	Get scheduling priority
sched_setparam()	SchedSet()	Set scheduling priority
sched_getscheduler()	SchedGet()	Get scheduling policy
sched_setscheduler()	SchedSet()	Set scheduling policy

IPC issues

Since all the threads in a process have unhindered access to the shared data space, wouldn't this execution model "trivially" solve all of our IPC problems? Can't we just communicate the data through shared memory and dispense with any other execution models and IPC mechanisms?

If only it were that simple!

One issue is that the access of individual threads to common data must be synchronized. Having one thread read inconsistent data because another thread is part way through modifying it is a recipe for disaster. For example, if one thread is updating a linked list, no other threads can be allowed to traverse or modify the list until the first thread has finished. A code passage that must execute "serially" (i.e. by only one thread at a time) in this manner is termed a "critical section." The program would fail (intermittently, depending on how frequently a "collision" occurred) with irreparably damaged links unless some synchronization mechanism ensured serial access.

Mutexes, semaphores, and condvars are examples of synchronization tools that can be used to address this problem. These tools are described later in this section.

Although synchronization services can be used to allow threads to cooperate, shared memory per se can't address a number of IPC issues. For example, although threads can communicate through the common data space, this works only if all the threads communicating are within a single process. What if our application needs to communicate a query to a database server? We need to pass the details of our query to the database server, but the thread we need to communicate with lies within a database server process and the address space of that server isn't addressable to us.

The OS takes care of the network-distributed IPC issue because the one interface -- message passing -- operates in both the local and network-remote cases, and can be used to access all OS services. Since messages can be exactly sized, and since most messages tend to be quite tiny (e.g. the error status on a write request, or a tiny read request), the data moved around the network can be far less with message passing than with network-distributed shared memory, which would tend to copy 4K pages around.

Thread complexity issues

Although threads are very appropriate for some system designs, it's important to respect the Pandora's box of complexities their use unleashes. In some ways, it's ironic that while MMU-protected multitasking has become common, computing fashion has made popular the use of multiple threads in an unprotected address space. This not only makes debugging difficult, but also hampers the generation of reliable code.

Threads were initially introduced to UNIX systems as a "light-weight" concurrency mechanism to address the problem of slow context switches between "heavy weight" processes. Although this is a worthwhile goal, an obvious question arises: Why are process-to-process context switches slow in the first place?

Architecturally, the OS addresses the context-switch performance issue first. In fact, threads and processes provide nearly identical context-switch performance numbers. QNX Neutrino's process-switch times are faster than UNIX thread-switch times. As a result, QNX Neutrino threads don't need to be used to solve the IPC performance problem; instead, they're a tool for achieving greater concurrency within application and server processes.

Without resorting to threads, fast process-to-process context switching makes it reasonable to structure an application as a team of cooperating processes sharing an explicitly allocated shared-memory region. An application thus exposes itself to bugs in the cooperating processes only so far as the effects of those bugs on the contents of the shared-memory region. The private memory of the process is still protected from the other processes. In the purely threaded model, the private data of all threads (including their stacks) is openly accessible, vulnerable to stray pointer errors in any thread in the process.

Nevertheless, threads can also provide concurrency advantages that a pure process model cannot address. For example, a filesystem server process that executes requests on behalf of many clients (where each request takes significant time to complete), definitely benefits from having multiple threads of execution. If one client process requests a block from disk, while another client requests a block already in cache, the filesystem process can utilize a pool of threads to concurrently service client requests, rather than remain "busy" until the disk block is read for the first request.

As requests arrive, each thread is able to respond directly from the buffer cache or to block and wait for disk I/O without increasing the response latency seen by other client processes. The filesystem server can "precreate" a team of threads, ready to respond in turn to client requests as they arrive. Although this complicates the architecture of the filesystem manager, the gains in concurrency are significant.

Synchronization services

QNX Neutrino provides the POSIX-standard thread-level synchronization primitives, some of which are useful even between threads in different processes. The synchronization services include at least the following:

Synchronization service	Supported between processes	Supported across a QNX LAN
Mutexes	Yes	No
Condvars	Yes	No
Barriers	No	No
Sleepon locks	No	No
Reader/writer locks	Yes	No
Semaphores	Yes	Yes (named only)
FIFO scheduling	Yes	No
Send/Receive/Reply	Yes	Yes
Atomic operations	Yes	No

Mutexes: mutual exclusion locks

Mutual exclusion locks, or mutexes, are the simplest of the synchronization services. A mutex is used to ensure exclusive access to data shared between threads. It is typically acquired (pthread_mutex_lock() or pthread_mutex_timedlock()) and released (pthread_mutex_unlock()) around the code that accesses the shared data (usually a critical section).

Only one thread may have the mutex locked at any given time. Threads attempting to lock an already locked mutex will block until the thread that owns the mutex unlocks it. When the thread unlocks the mutex, the highest-priority thread waiting to lock the mutex will unblock and become the new owner of the mutex. In this way, threads will sequence through a critical region in priority-order.

On most processors, acquisition of a mutex doesn't require entry to the kernel for a free mutex. What allows this is the use of the compare-and-swap opcode on x86 processors and the load/store conditional opcodes on most RISC processors.

Entry to the kernel is done at acquisition time only if the mutex is already held so that the thread can go on a blocked list; kernel entry is done on exit if other threads are waiting to be unblocked on that mutex. This allows acquisition and release of an uncontested critical section or resource to be very quick, incurring work by the OS only to resolve contention.

A nonblocking lock function (pthread_mutex_trylock()) can be used to test whether the mutex is currently locked or not. For best performance, the execution time of the critical section should be small and of bounded duration. A condvar should be used if the thread may block within the critical section.

Priority inheritance and mutexes

By default, if a thread with a higher priority than the mutex owner attempts to lock a mutex, then the effective priority of the current owner is increased to that of the higher-priority blocked thread waiting for the mutex. The current owner returns to its real priority when it unlocks the mutex. This scheme not only ensures that the higher-priority thread will be blocked waiting for the mutex for the shortest possible time, but also solves the classic priority-inversion problem.

The pthread_mutexattr_init() function sets the protocol to PTHREAD_PRIO_INHERIT to allow this behavior; you can call pthread_mutexattr_setprotocol() to override this setting. The pthread_mutex_trylock() function doesn't change the thread priorities because it doesn't block.

You can also modify the attributes of the mutex (using pthread_mutexattr_setrecursive()) to allow a mutex to be recursively locked by the same thread. This can be useful to allow a thread to call a routine that might attempt to lock a mutex that the thread already happens to have locked.

Condvars: condition variables

A condition variable, or condvar, is used to block a thread within a critical section until some condition is satisfied. The condition can be arbitrarily complex and is independent of the condvar. However, the condvar must always be used with a mutex lock in order to implement a monitor.

A condvar supports three operations:

• wait (pthread_cond_wait())
• signal (pthread_cond_signal())
• broadcast (pthread_cond_broadcast()).

Here's a typical example of how a condvar can be used:

pthread_mutex_lock( &m );
. . .
while (!arbitrary_condition) {
    pthread_cond_wait( &cv, &m );
    }
. . .
pthread_mutex_unlock( &m );

In this code sample, the mutex is acquired before the condition is tested. This ensures that only this thread has access to the arbitrary condition being examined. While the condition is true, the code sample will block on the wait call until some other thread performs a signal or broadcast on the condvar.

The while loop is required for two reasons. First of all, POSIX cannot guarantee that false wakeups will not occur (e.g. multiprocessor systems). Second, when another thread has made a modification to the condition, we need to retest to ensure that the modification matches our criteria. The associated mutex is unlocked atomically by pthread_cond_wait() when the waiting thread is blocked to allow another thread to enter the critical section.

A thread that performs a signal will unblock the highest-priority thread queued on the condvar, while a broadcast will unblock all threads queued on the condvar. The associated mutex is locked atomically by the highest-priority unblocked thread; the thread must then unlock the mutex after proceeding through the critical section.

A version of the condvar wait operation allows a timeout to be specified (pthread_cond_timedwait()). The waiting thread can then be unblocked when the timeout expires.

Barriers

A barrier is a synchronization mechanism that lets you "corral" several cooperating threads (e.g. in a matrix computation), forcing them to wait at a specific point until all have finished before any one thread can continue.

Unlike the pthread_join() function, where you'd wait for the threads to terminate, in the case of a barrier you're waiting for the threads to rendezvous at a certain point. When the specified number of threads arrive at the barrier, we unblock all of them so they can continue to run.

You first create a barrier with pthread_barrier_init():

#include <pthread.h>

int
pthread_barrier_init (pthread_barrier_t *barrier,
                      const pthread_barrierattr_t *attr,
                      unsigned int count);

This creates a barrier object at the passed address (a pointer to the barrier object is in barrier), with the attributes as specified by attr. The count member holds the number of threads that must call pthread_barrier_wait().

Once the barrier is created, each thread will call pthread_barrier_wait() to indicate that it has completed:

#include <pthread.h>

int pthread_barrier_wait (pthread_barrier_t *barrier);

When a thread calls pthread_barrier_wait(), it blocks until the number of threads specified initially in the pthread_barrier_init() function have called pthread_barrier_wait() (and blocked also). When the correct number of threads have called pthread_barrier_wait(), all those threads will unblock at the same time.

Here's an example:

/*
 *  barrier1.c
*/

#include <stdio.h>
#include <time.h>
#include <pthread.h>
#include <sys/neutrino.h>

pthread_barrier_t   barrier; // barrier synchronization object

main () // ignore arguments
{
    time_t  now;

    // create a barrier object with a count of 3
    pthread_barrier_init (&barrier, NULL, 3);

    // start up two threads, thread1 and thread2
    pthread_create (NULL, NULL, thread1, NULL);
    pthread_create (NULL, NULL, thread2, NULL);

    // at this point, thread1 and thread2 are running

    // now wait for completion
    time (&now);
    printf ("main() waiting for barrier at %s", ctime (&now));
    pthread_barrier_wait (&barrier);

    // after this point, all three threads have completed.
    time (&now);
    printf ("barrier in main() done at %s", ctime (&now));
}

void *
thread1 (void *not_used)
{
    time_t  now;

    time (&now);
    printf ("thread1 starting at %s", ctime (&now));

    // do the computation
    // let's just do a sleep here...
    sleep (20);
    pthread_barrier_wait (&barrier);
    // after this point, all three threads have completed.
    time (&now);
    printf ("barrier in thread1() done at %s", ctime (&now));
}

void *
thread2 (void *not_used)
{
    time_t  now;

    time (&now);
    printf ("thread2 starting at %s", ctime (&now));

    // do the computation
    // let's just do a sleep here...
    sleep (40);
    pthread_barrier_wait (&barrier);
    // after this point, all three threads have completed.
    time (&now);
    printf ("barrier in thread2() done at %s", ctime (&now));
}

The main thread created the barrier object and initialized it with a count of the total number of threads that must be synchronized to the barrier before the threads may carry on. In the example above, we used a count of 3: one for the main() thread, one for thread1(), and one for thread2().

Then we start thread1() and thread2(). To simplify this example, we have the threads sleep to cause a delay, as if computations were occurring. To synchronize, the main thread simply blocks itself on the barrier, knowing that the barrier will unblock only after the two worker threads have joined it as well.

In this release, the following barrier functions are included:

Function	Description
pthread_barrierattr_getpshared()	Get the value of a barrier's process-shared attribute
pthread_barrierattr_destroy()	Destroy a barrier's attributes object
pthread_barrierattr_init()	Initialize a barrier's attributes object
pthread_barrierattr_setpshared()	Set the value of a barrier's process-shared attribute
pthread_barrier_destroy()	Destroy a barrier
pthread_barrier_init()	Initialize a barrier
pthread_barrier_wait()	Synchronize participating threads at the barrier

Sleepon locks

Sleepon locks are very similar to condvars, with a few subtle differences. Like condvars, sleepon locks (pthread_sleepon_lock()) can be used to block until a condition becomes true (like a memory location changing value). But unlike condvars, which must be allocated for each condition to be checked, sleepon locks multiplex their functionality over a single mutex and dynamically allocated condvar, regardless of the number of conditions being checked. The maximum number of condvars ends up being equal to the maximum number of blocked threads. These locks are patterned after the sleepon locks commonly used within the UNIX kernel.

Reader/writer locks

More formally known as "Multiple readers, single writer locks," these locks are used when the access pattern for a data structure consists of many threads reading the data, and (at most) one thread writing the data. These locks are more expensive than mutexes, but can be useful for this data access pattern.

This lock works by allowing all the threads that request a read-access lock (pthread_rwlock_rdlock()) to succeed in their request. But when a thread wishing to write asks for the lock (pthread_rwlock_wrlock()), the request is denied until all the current reading threads release their reading locks (pthread_rwlock_unlock()).

Multiple writing threads can queue (in priority order) waiting for their chance to write the protected data structure, and all the blocked writer-threads will get to run before reading threads are allowed access again. The priorities of the reading threads are not considered.

There are also calls (pthread_rwlock_tryrdlock() and pthread_rwlock_trywrlock()) to allow a thread to test the attempt to achieve the requested lock, without blocking. These calls return with a successful lock or a status indicating that the lock couldn't be granted immediately.

Reader/writer locks aren't implemented directly within the kernel, but are instead built from the mutex and condvar services provided by the kernel.

Semaphores

Semaphores are another common form of synchronization that allows threads to "post" (sem_post()) and "wait" (sem_wait()) on a semaphore to control when threads wake or sleep. The post operation increments the semaphore; the wait operation decrements it.

If you wait on a semaphore that is positive, you will not block. Waiting on a nonpositive semaphore will block until some other thread executes a post. It is valid to post one or more times before a wait. This use will allow one or more threads to execute the wait without blocking.

A significant difference between semaphores and other synchronization primitives is that semaphores are "async safe" and can be manipulated by signal handlers. If the desired effect is to have a signal handler wake a thread, semaphores are the right choice.

Another useful property of semaphores is that they were defined to operate between processes. Although our mutexes work between processes, the POSIX thread standard considers this an optional capability and as such may not be portable across systems. For synchronization between threads in a single process, mutexes will be more efficient than semaphores.

As a useful variation, a named semaphore service is also available. It lets you use semaphores between processes on different machines connected by a network.

Since semaphores, like condition variables, can legally return a nonzero value because of a false wakeup, correct usage requires a loop:

while (sem_wait(&s) && (errno == EINTR)) { do_nothing(); }
do_critical_region();   /* Semaphore was decremented */

Synchronization via scheduling algorithm

By selecting the POSIX FIFO scheduling algorithm, we can guarantee that no two threads of the same priority execute the critical section concurrently on a non-SMP system. The FIFO scheduling algorithm dictates that all FIFO-scheduled threads in the system at the same priority will run, when scheduled, until they voluntarily release the processor to another thread.

This "release" can also occur when the thread blocks as part of requesting the service of another process, or when a signal occurs. The critical region must therefore be carefully coded and documented so that later maintenance of the code doesn't violate this condition.

In addition, higher-priority threads in that (or any other) process could still preempt these FIFO-scheduled threads. So, all the threads that could "collide" within the critical section must be FIFO-scheduled at the same priority. Having enforced this condition, the threads can then casually access this shared memory without having to first make explicit synchronization calls.

Synchronization via message passing

Our Send/Receive/Reply message-passing IPC services (described later) implement an implicit synchronization by their blocking nature. These IPC services can, in many instances, render other synchronization services unnecessary. They are also the only synchronization and IPC primitives (other than named semaphores, which are built on top of messaging) that can be used across the network.

Synchronization via atomic operations

In some cases, you may want to perform a short operation (such as incrementing a variable) with the guarantee that the operation will perform atomically -- i.e. the operation won't be preempted by another thread or ISR (Interrupt Service Routine).

Under QNX Neutrino, we provide atomic operations for:

• adding a value
• subtracting a value
• clearing bits
• setting bits
• toggling (complementing) bits.

These atomic operations are available by including the C header file <atomic.h>.

Although you can use these atomic operations just about anywhere, you'll find them particularly useful in these two cases:

• between an ISR and a thread
• between two threads (SMP or single-processor).

Since an ISR can preempt a thread at any given point, the only way that the thread would be able to protect itself would be to disable interrupts. Since you should avoid disabling interrupts in a realtime system, we recommend that you use the atomic operations provided with QNX Neutrino.

On an SMP system, multiple threads can and do run concurrently. Again, we run into the same situation as with interrupts above -- you should use the atomic operations where applicable to eliminate the need to disable and reenable interrupts.

Synchronization services implementation

The following table lists the various microkernel calls and the higher-level POSIX calls constructed from them:

Microkernel call	POSIX call	Description
SyncTypeCreate()	pthread_mutex_init(), pthread_cond_init(), sem_init()	Create object for mutex, condvars, and semaphore.
SyncDestroy()	pthread_mutex_destroy(), pthread_cond_destroy(), sem_destroy()	Destroy synchronization object.
SyncCondvarWait()	pthread_cond_wait(), pthread_cond_timedwait()	Block on a condvar.
SyncCondvarSignal()	pthread_cond_broadcast(), pthread_cond_signal()	Wake up condvar-blocked threads.
SyncMutexLock()	pthread_mutex_lock(), pthread_mutex_trylock()	Lock a mutex.
SyncMutexUnlock()	pthread_mutex_unlock()	Unlock a mutex.
SyncSemPost()	sem_post()	Post a semaphore.
SyncSemWait()	sem_wait(), sem_trywait()	Wait on a semaphore.

QNX Neutrino IPC

IPC plays a fundamental role in the transformation of QNX Neutrino from an embedded realtime kernel into a full-scale POSIX operating system. As various service-providing processes are added to the microkernel, IPC is the "glue" that connects those components into a cohesive whole.

Although message passing is the primary form of IPC in QNX Neutrino, several other forms are available as well. Unless otherwise noted, those other forms of IPC are built over our native message passing. The strategy is to create a simple, robust IPC service that can be tuned for performance through a simplified code path in the microkernel; more "feature cluttered" IPC services can then be implemented from these.

Benchmarks comparing higher-level IPC services (like pipes and FIFOs implemented over our messaging) with their monolithic kernel counterparts show comparable performance.

QNX Neutrino offers at least the following forms of IPC:

Service:	Implemented in:
Message-passing	Kernel
Signals	Kernel
POSIX message queues	External process
Shared memory	Process manager
Pipes	External process
FIFOs	External process

The designer can select these services on the basis of bandwidth requirements, the need for queuing, network transparency, etc. The tradeoff can be complex, but the flexibility is useful.

As part of the engineering effort that went into defining the QNX Neutrino microkernel, the focus on message passing as the fundamental IPC primitive was deliberate. As a form of IPC, message passing (as implemented in MsgSend(), MsgReceive(), and MsgReply()), is synchronous and copies data. Let's explore these two attributes in more detail.

Synchronous message passing

A thread that does a MsgSend() to another thread (which could be within another process) will be blocked until the target thread does a MsgReceive(), processes the message, and executes a MsgReply(). If a thread executes a MsgReceive() without a previously sent message pending, it will block until another thread executes a MsgSend().

In Neutrino, a server thread typically loops, waiting to receive a message from a client thread. As described earlier, a thread -- whether a server or a client -- is in the READY state if it can use the CPU. It might not actually be getting any CPU time because of its and other threads' priority and scheduling algorithm, but the thread isn't blocked.

Let's look first at the client thread:

---

---

Changes of state for a client thread in a send-receive-reply transaction.

• If the client thread calls MsgSend(), and the server thread hasn't yet called MsgReceive(), then the client thread becomes SEND blocked. Once the server thread calls MsgReceive(), the kernel changes the client thread's state to be REPLY blocked, which means that server thread has received the message and now must reply. When the server thread calls MsgReply(), the client thread becomes READY.
• If the client thread calls MsgSend(), and the server thread is already blocked on the MsgReceive(), then the client thread immediately becomes REPLY blocked, skipping the SEND-blocked state completely.
• If the server thread fails, exits, or disappears, the client thread becomes READY, with MsgSend() indicating an error.

Next, let's consider the server thread:

---

---

Changes of state for a server thread in a send-receive-reply transaction.

• If the server thread calls MsgReceive(), and no other thread has sent to it, then the server thread becomes RECEIVE blocked. When another thread sends to it, the server thread becomes READY.
• If the server thread calls MsgReceive(), and another thread has already sent to it, then MsgReceive() returns immediately with the message. In this case, the server thread doesn't block.
• If the server thread calls MsgReply(), it doesn't become blocked.

This inherent blocking synchronizes the execution of the sending thread, since the act of requesting that the data be sent also causes the sending thread to be blocked and the receiving thread to be scheduled for execution. This happens without requiring explicit work by the kernel to determine which thread to run next (as would be the case with most other forms of IPC). Execution and data move directly from one context to another.

Data-queuing capabilities are omitted from these messaging primitives because queueing could be implemented when needed within the receiving thread. The sending thread is often prepared to wait for a response; queueing is unnecessary overhead and complexity (i.e. it slows down the nonqueued case). As a result, the sending thread doesn't need to make a separate, explicit blocking call to wait for a response (as it would if some other IPC form had been used).

While the send and receive operations are blocking and synchronous, MsgReply() (or MsgError()) doesn't block. Since the client thread is already blocked waiting for the reply, no additional synchronization is required, so a blocking MsgReply() isn't needed. This allows a server to reply to a client and continue processing while the kernel and/or networking code asynchronously passes the reply data to the sending thread and marks it ready for execution. Since most servers will tend to do some processing to prepare to receive the next request (at which point they block again), this works out well.

MsgReply() vs MsgError()

The MsgReply() function is used to return a status and zero or more bytes to the client. MsgError(), on the other hand, is used to return only a status to the client. Both functions will unblock the client from its MsgSend().

Message copying

Since our messaging services copy a message directly from the address space of one thread to another without intermediate buffering, the message-delivery performance approaches the memory bandwidth of the underlying hardware. The kernel attaches no special meaning to the content of a message -- the data in a message has meaning only as mutually defined by sender and receiver. However, "well-defined" message types are also provided so that user-written processes or threads can augment or substitute for system-supplied services.

The messaging primitives support multipart transfers, so that a message delivered from the address space of one thread to another needn't pre-exist in a single, contiguous buffer. Instead, both the sending and receiving threads can specify a vector table that indicates where the sending and receiving message fragments reside in memory. Note that the size of the various parts can be different for the sender and receiver.

Multipart transfers allow messages that have a header block separate from the data block to be sent without performance-consuming copying of the data to create a contiguous message. In addition, if the underlying data structure is a ring buffer, specifying a three-part message will allow a header and two disjoint ranges within the ring buffer to be sent as a single atomic message. A hardware equivalent of this concept would be that of a scatter/gather DMA facility.

---

---

A multipart transfer.

The multipart transfers are also used extensively by filesystems. On a read, the data is copied directly from the filesystem cache into the application using a message with n parts for the data. Each part points into the cache and compensates for the fact that cache blocks aren't contiguous in memory with a read starting or ending within a block.

For example, with a cache block size of 512 bytes, a read of 1454 bytes can be satisfied with a 5-part message:

---

---

Scatter/gather of a read of 1454 bytes.

Since message data is explicitly copied between address spaces (rather than by doing page table manipulations), messages can be easily allocated on the stack instead of from a special pool of page-aligned memory for MMU "page flipping." As a result, many of the library routines that implement the API between client and server processes can be trivially expressed, without elaborate IPC-specific memory allocation calls.

For example, the code used by a client thread to request that the filesystem manager execute lseek on its behalf is implemented as follows:

#include <unistd.h>
#include <errno.h>
#include <sys/iomsg.h>

off64_t lseek64(int fd, off64_t offset, int whence) {
    io_lseek_t                            msg;
    off64_t                               off;

    msg.i.type = _IO_LSEEK;
    msg.i.combine_len = sizeof msg.i;
    msg.i.offset = offset;
    msg.i.whence = whence;
    msg.i.zero = 0;
    if(MsgSend(fd, &msg.i, sizeof msg.i, &off, sizeof off) == -1) {
        return -1;
    }
    return off;
}

off64_t tell64(int fd) {
    return lseek64(fd, 0, SEEK_CUR);
}

off_t lseek(int fd, off_t offset, int whence) {
    return lseek64(fd, offset, whence);
}

off_t tell(int fd) {
    return lseek64(fd, 0, SEEK_CUR);
}

This code essentially builds a message structure on the stack, populates it with various constants and passed parameters from the calling thread, and sends it to the filesystem manager associated with fd. The reply indicates the success or failure of the operation.

Simple messages

For simple single-part messages, the OS provides functions that take a pointer directly to a buffer without the need for an IOV (input/output vector). In this case, the number of parts is replaced by the size of the message directly pointed to. In the case of the message send primitive -- which takes a send and a reply buffer -- this introduces four variations:

Function	Send message	Reply message
MsgSend()	Simple	Simple
MsgSendsv()	Simple	IOV
MsgSendvs()	IOV	Simple
MsgSendv()	IOV	IOV

The other messaging primitives that take a direct message simply drop the trailing "v" in their names:

IOV	Simple direct
MsgReceivev()	MsgReceive()
MsgReceivePulsev()	MsgReceivePulse()
MsgReplyv()	MsgReply()
MsgReadv()	MsgRead()
MsgWritev()	MsgWrite()

Channels and connections

In QNX Neutrino, message passing is directed towards channels and connections, rather than targeted directly from thread to thread. A thread that wishes to receive messages first creates a channel; another thread that wishes to send a message to that thread must first make a connection by "attaching" to that channel.

Channels are required by the message kernel calls and are used by servers to MsgReceive() messages on. Connections are created by client threads to "connect" to the channels made available by servers. Once connections are established, clients can MsgSend() messages over them. If a number of threads in a process all attach to the same channel, then the connections all map to the same kernel object for efficiency. Channels and connections are named within a process by a small integer identifier. Client connections map directly into file descriptors.

Architecturally, this is a key point. By having client connections map directly into FDs, we have eliminated yet another layer of translation. We don't need to "figure out" where to send a message based on the file descriptor (e.g. via a read(fd) call). Instead, we can simply send a message directly to the "file descriptor" (i.e. connection ID).

Function	Description
ChannelCreate()	Create a channel to receive messages on.
ChannelDestroy()	Destroy a channel.
ConnectAttach()	Create a connection to send messages on.
ConnectDetach()	Detach a connection.

---

---

Connections map elegantly into file descriptors.

A process acting as a server would implement an event loop to receive and process messages as follows:

chid = ChannelCreate(flags);
SETIOV(&iov, &msg, sizeof(msg));
for(;;) {
    rcv_id = MsgReceivev( chid, &iov, parts, &info );

    switch( msg.type ) {
        /* Perform message processing here */
        }

    MsgReplyv( rcv_id, &iov, rparts );
    }

This loop allows the thread to receive messages from any thread that had a connection to the channel.

The channel has several lists of messages associated with it:

Receive

A LIFO queue of threads waiting for messages.

Send

A priority FIFO queue of threads that have sent messages that haven't yet been received.

Reply

An unordered list of threads that have sent messages that have been received, but not yet replied to.

While in any of these lists, the waiting thread is blocked (i.e. RECEIVE-, SEND-, or REPLY-blocked). Multiple threads and multiple clients may wait on one channel.

Pulses

In addition to the synchronous Send/Receive/Reply services, the OS also supports fixed-size, nonblocking messages. These are referred to as pulses and carry a small payload (four bytes of data plus a single byte code).

Pulses pack a relatively small payload -- eight bits of code and 32 bits of data. Pulses are often used as a notification mechanism within interrupt handlers. They also allow servers to signal clients without blocking on them.

---

---

Pulses pack a small payload.

Priority inheritance and messages

A server process receives messages and pulses in priority order. As the threads within the server receive requests, they then inherit the priority (but not the scheduling algorithm) of the sending thread. As a result, the relative priorities of the threads requesting work of the server are preserved, and the server work will be executed at the appropriate priority. This message-driven priority inheritance avoids priority-inversion problems.

For example, suppose the system includes the following:

• a server thread, running at priority 22
• a client thread, T1, running at priority 13
• a client thread, T2, running at priority 10

Without priority inheritance, if T2 sends a message to the server, it's effectively getting work done for it at priority 22, so T2's priority has been inverted.

What actually happens is that when the server receives a message, its effective priority changes to that of the highest-priority sender. In this case, T2's priority is lower than the server's, so the change in the server's effective priority takes place when the server receives the message.

Next, suppose that T1 sends a message to the server while it's still at priority 10. Since T1's priority is higher than the server's current priority, the change in the server's priority happens when T1 sends the message.

The change happens before the server receives the message to avoid another case of priority inversion. If the server's priority remains unchanged at 10, and another thread, T3, starts to run at priority 11, the server has to wait until T3 lets it have some CPU time so that it can eventually receive T1's message. So, T1 would would be delayed by a lower-priority thread, T3.

You can turn off priority inheritance by specifying the _NTO_CHF_FIXED_PRIORITY flag when you call ChannelCreate(). If you're using adaptive partitioning, this flag also causes the receiving threads not to run in the sending threads' partitions.

Message-passing API

The message-passing API consists of the following functions:

Function	Description
MsgSend()	Send a message and block until reply.
MsgReceive()	Wait for a message.
MsgReceivePulse()	Wait for a tiny, nonblocking message (pulse).
MsgReply()	Reply to a message.
MsgError()	Reply only with an error status. No message bytes are transferred.
MsgRead()	Read additional data from a received message.
MsgWrite()	Write additional data to a reply message.
MsgInfo()	Obtain info on a received message.
MsgSendPulse()	Send a tiny, nonblocking message (pulse).
MsgDeliverEvent()	Deliver an event to a client.
MsgKeyData()	Key a message to allow security checks.

Robust implementations with Send/Receive/Reply

Architecting a QNX Neutrino application as a team of cooperating threads and processes via Send/Receive/Reply results in a system that uses synchronous notification. IPC thus occurs at specified transitions within the system, rather than asynchronously.

A significant problem with asynchronous systems is that event notification requires signal handlers to be run. Asynchronous IPC can make it difficult to thoroughly test the operation of the system and make sure that no matter when the signal handler runs, that processing will continue as intended. Applications often try to avoid this scenario by relying on a "window" explicitly opened and shut, during which signals will be tolerated.

With a synchronous, nonqueued system architecture built around Send/Receive/Reply, robust application architectures can be very readily implemented and delivered.

Avoiding deadlock situations is another difficult problem when constructing applications from various combinations of queued IPC, shared memory, and miscellaneous synchronization primitives. For example, suppose thread A doesn't release mutex 1 until thread B releases mutex 2. Unfortunately, if thread B is in the state of not releasing mutex 2 until thread A releases mutex 1, a standoff results. Simulation tools are often invoked in order to ensure that deadlock won't occur as the system runs.

The Send/Receive/Reply IPC primitives allow the construction of deadlock-free systems with the observation of only these simple rules:

1. Never have two threads send to each other.
2. Always arrange your threads in a hierarchy, with sends going up the tree.

The first rule is an obvious avoidance of the standoff situation, but the second rule requires further explanation. The team of cooperating threads and processes is arranged as follows:

---

---

Threads should always send up to higher-level threads.

Here the threads at any given level in the hierarchy never send to each other, but send only upwards instead.

One example of this might be a client application that sends to a database server process, which in turn sends to a filesystem process. Since the sending threads block and wait for the target thread to reply, and since the target thread isn't send-blocked on the sending thread, deadlock can't happen.

But how does a higher-level thread notify a lower-level thread that it has the results of a previously requested operation? (Assume the lower-level thread didn't want to wait for the replied results when it last sent.)

QNX Neutrino provides a very flexible architecture with the MsgDeliverEvent() kernel call to deliver nonblocking events. All of the common asynchronous services can be implemented with this. For example, the server-side of the select() call is an API that an application can use to allow a thread to wait for an I/O event to complete on a set of file descriptors. In addition to an asynchronous notification mechanism being needed as a "back channel" for notifications from higher-level threads to lower-level threads, we can also build a reliable notification system for timers, hardware interrupts, and other event sources around this.

---

---

A higher-level thread can "send" a pulse event.

A related issue is the problem of how a higher-level thread can request work of a lower-level thread without sending to it, risking deadlock. The lower-level thread is present only to serve as a "worker thread" for the higher-level thread, doing work on request. The lower-level thread would send in order to "report for work," but the higher-level thread wouldn't reply then. It would defer the reply until the higher-level thread had work to be done, and it would reply (which is a nonblocking operation) with the data describing the work. In effect, the reply is being used to initiate work, not the send, which neatly side-steps rule #1.

Events

A significant advance in the kernel design for QNX Neutrino is the event-handling subsystem. POSIX and its realtime extensions define a number of asynchronous notification methods (e.g. UNIX signals that don't queue or pass data, POSIX realtime signals that may queue and pass data, etc.).

The kernel also defines additional, QNX-specific notification techniques such as pulses. Implementing all of these event mechanisms could have consumed significant code space, so our implementation strategy was to build all of these notification methods over a single, rich, event subsystem.

A benefit of this approach is that capabilities exclusive to one notification technique can become available to others. For example, an application can apply the same queueing services of POSIX realtime signals to UNIX signals. This can simplify the robust implementation of signal handlers within applications.

The events encountered by an executing thread can come from any of three sources:

• a MsgDeliverEvent() kernel call invoked by a thread
• an interrupt handler
• the expiry of a timer.

The event itself can be any of a number of different types: QNX Neutrino pulses, interrupts, various forms of signals, and forced "unblock" events. "Unblock" is a means by which a thread can be released from a deliberately blocked state without any explicit event actually being delivered.

Given this multiplicity of event types, and applications needing the ability to request whichever asynchronous notification technique best suits their needs, it would be awkward to require that server processes (the higher-level threads from the previous section) carry code to support all these options.

Instead, the client thread can give a data structure, or "cookie," to the server to hang on to until later. When the server needs to notify the client thread, it will invoke MsgDeliverEvent() and the microkernel will set the event type encoded within the cookie upon the client thread.

---

---

The client sends a sigevent to the server.

I/O notification

The ionotify() function is a means by which a client thread can request asynchronous event delivery. Many of the POSIX asynchronous services (e.g. mq_notify() and the client-side of the select()) are built on top of it. When performing I/O on a file descriptor (fd), the thread may choose to wait for an I/O event to complete (for the write() case), or for data to arrive (for the read() case). Rather than have the thread block on the resource manager process that's servicing the read/write request, ionotify() can allow the client thread to post an event to the resource manager that the client thread would like to receive when the indicated I/O condition occurs. Waiting in this manner allows the thread to continue executing and responding to event sources other than just the single I/O request.

The select() call is implemented using I/O notification and allows a thread to block and wait for a mix of I/O events on multiple fd's while continuing to respond to other forms of IPC.

Here are the conditions upon which the requested event can be delivered:

• _NOTIFY_COND_OUTPUT -- there's room in the output buffer for more data.
• _NOTIFY_COND_INPUT -- resource-manager-defined amount of data is available to read.
• _NOTIFY_OUT_OF_BAND -- resource-manager-defined "out of band" data is available.

Signals

The OS supports the 32 standard POSIX signals (as in UNIX) as well as the POSIX realtime signals, both numbered from a kernel-implemented set of 64 signals with uniform functionality. While the POSIX standard defines realtime signals as differing from UNIX-style signals (in that they may contain four bytes of data and a byte code and may be queued for delivery), this functionality can be explicitly selected or deselected on a per-signal basis, allowing this converged implementation to still comply with the standard.

Incidentally, the UNIX-style signals can select POSIX realtime signal queuing, if the application wants it. QNX Neutrino also extends the signal-delivery mechanisms of POSIX by allowing signals to be targeted at specific threads, rather than simply at the process containing the threads. Since signals are an asynchronous event, they're also implemented with the event-delivery mechanisms.

Microkernel call	POSIX call	Description
SignalKill()	kill(), pthread_kill(), raise(), sigqueue()	Set a signal on a process group, process, or thread.
SignalAction()	sigaction()	Define action to take on receipt of a signal.
SignalProcmask()	sigprocmask(), pthread_sigmask()	Change signal blocked mask of a thread.
SignalSuspend()	sigsuspend(), pause()	Block until a signal invokes a signal handler.
SignalWaitinfo()	sigwaitinfo()	Wait for signal and return info on it.

The original POSIX specification defined signal operation on processes only. In a multithreaded process, the following rules are followed:

• The signal actions are maintained at the process level. If a thread ignores or catches a signal, it affects all threads within the process.
• The signal mask is maintained at the thread level. If a thread blocks a signal, it affects only that thread.
• An un-ignored signal targeted at a thread will be delivered to that thread alone.
• An un-ignored signal targeted at a process is delivered to the first thread that doesn't have the signal blocked. If all threads have the signal blocked, the signal will be queued on the process until any thread ignores or unblocks the signal. If ignored, the signal on the process will be removed. If unblocked, the signal will be moved from the process to the thread that unblocked it.

When a signal is targeted at a process with a large number of threads, the thread table must be scanned, looking for a thread with the signal unblocked. Standard practice for most multithreaded processes is to mask the signal in all threads but one, which is dedicated to handling them. To increase the efficiency of process-signal delivery, the kernel will cache the last thread that accepted a signal and will always attempt to deliver the signal to it first.

---

---

Signal delivery.

The POSIX standard includes the concept of queued realtime signals. QNX Neutrino supports optional queuing of any signal, not just realtime signals. The queuing can be specified on a signal-by-signal basis within a process. Each signal can have an associated 8-bit code and a 32-bit value.

This is very similar to message pulses described earlier. The kernel takes advantage of this similarity and uses common code for managing both signals and pulses. The signal number is mapped to a pulse priority using _SIGMAX - signo. As a result, signals are delivered in priority order with lower signal numbers having higher priority. This conforms with the POSIX standard, which states that existing signals have priority over the new realtime signals.

Special signals

As mentioned earlier, the OS defines a total of 64 signals. Their range is as follows:

Signal range	Description
1 ... 57	57 POSIX signals (including traditional UNIX signals)
41 ... 56	16 POSIX realtime signals (SIGRTMIN to SIGRTMAX)
57 ... 64	Eight special-purpose QNX Neutrino signals

The eight special signals cannot be ignored or caught. An attempt to call the signal() or sigaction() functions or the SignalAction() kernel call to change them will fail with an error of EINVAL.

In addition, these signals are always blocked and have signal queuing enabled. An attempt to unblock these signals via the sigprocmask() function or SignalProcmask() kernel call will be quietly ignored.

A regular signal can be programmed to this behavior using the following standard signal calls. The special signals save the programmer from writing this code and protect the signal from accidental changes to this behavior.

sigset_t *set;
struct sigaction action;

sigemptyset(&set);
sigaddset(&set, signo);
sigprocmask(SIG_BLOCK, &set, NULL);

action.sa_handler = SIG_DFL;
action.sa_flags = SA_SIGINFO;
sigaction(signo, &action, NULL);

This configuration makes these signals suitable for synchronous notification using the sigwaitinfo() function or SignalWaitinfo() kernel call. The following code will block until the eighth special signal is received:

sigset_t *set;
siginfo_t info;

sigemptyset(&set);
sigaddset(&set, SIGRTMAX + 8);
sigwaitinfo(&set, &info);
printf("Received signal %d with code %d and value %d\n",
            info.si_signo,
            info.si_code,
            info.si_value.sival_int);

Since the signals are always blocked, the program cannot be interrupted or killed if the special signal is delivered outside of the sigwaitinfo() function. Since signal queuing is always enabled, signals won't be lost -- they'll be queued for the next sigwaitinfo() call.

These signals were designed to solve a common IPC requirement where a server wishes to notify a client that it has information available for the client. The server will use the MsgDeliverEvent() call to notify the client. There are two reasonable choices for the event within the notification: pulses or signals.

A pulse is the preferred method for a client that may also be a server to other clients. In this case, the client will have created a channel for receiving messages and can also receive the pulse.

This won't be true for most simple clients. In order to receive a pulse, a simple client would be forced to create a channel for this express purpose. A signal can be used in place of a pulse if the signal is configured to be synchronous (i.e. the signal is blocked) and queued -- this is exactly how the special signals are configured. The client would replace the MsgReceive() call used to wait for a pulse on a channel with a simple sigwaitinfo() call to wait for the signal.

This signal mechanism is used by Photon to wait for events and by the select() function to wait for I/O from multiple servers. Of the eight special signals, the first two have been given special names for this use.

#define SIGSELECT   (SIGRTMAX + 1)
#define SIGPHOTON   (SIGRTMAX + 2)

Summary of signals

Signal	Description
SIGABRT	Abnormal termination signal such as issued by the abort() function.
SIGALRM	Timeout signal such as issued by the alarm() function.
SIGBUS	Indicates a memory parity error (QNX-specific interpretation). Note that if a second fault occurs while your process is in a signal handler for this fault, the process will be terminated.
SIGCHLD	Child process terminated. The default action is to ignore the signal.
SIGCONT	Continue if HELD. The default action is to ignore the signal if the process isn't HELD.
SIGDEADLK	Mutex deadlock occurred. If you haven't called SyncMutexEvent(), and if the conditions that would cause the kernel to deliver the event occur, then the kernel delivers a SIGDEADLK instead.
SIGEMT	EMT instruction (emulator trap).
SIGFPE	Erroneous arithmetic operation (integer or floating point), such as division by zero or an operation resulting in overflow. Note that if a second fault occurs while your process is in a signal handler for this fault, the process will be terminated.
SIGHUP	Death of session leader, or hangup detected on controlling terminal.
SIGILL	Detection of an invalid hardware instruction. Note that if a second fault occurs while your process is in a signal handler for this fault, the process will be terminated.
SIGINT	Interactive attention signal (Break).
SIGIOT	IOT instruction (not generated on x86 hardware).
SIGKILL	Termination signal -- should be used only for emergency situations. This signal cannot be caught or ignored.
SIGPIPE	Attempt to write on a pipe with no readers.
SIGPOLL	Pollable event occurred.
SIGQUIT	Interactive termination signal.
SIGSEGV	Detection of an invalid memory reference. Note that if a second fault occurs while your process is in a signal handler for this fault, the process will be terminated.
SIGSTOP	Stop process (the default). This signal cannot be caught or ignored.
SIGSYS	Bad argument to system call.
SIGTERM	Termination s